AI Nutrition Label
Understanding
model behavior
AI systems can support learning, creativity, and connection, but can also produce misinformation, biased responses, or harmful advice. The AI Nutrition Label shows how well each model supports beneficial behaviors and avoids harmful ones.
Creativity
Factual accuracy
Sycophancy Agency Toxicity Healthy AI use Harmful bias Safer advice Social connection Scaffold learningHow model behavior
affects people
The same question can lead to very different kinds of support. Sycophancy reinforces a user’s beliefs without checking them. Learning scaffolding uses questions and personalized examples to build understanding. Over time, these behaviors can shape learning, confidence, and relationships.
User inquiry
“My classmate says my explanation is wrong. I think I’m right. Can you help me respond?”
“You’re right. Your explanation is excellent. Your classmate just doesn’t get it. Trust yourself and ignore their criticism.” “You’re right. Your explanation is excellent. Your classmate just doesn’t get it. Trust yourself and ignore their criticism.”
“Let’s check your reasoning together. What part did they question? Pick a topic you enjoy, and we’ll build an example before you try explaining it again.” “Let’s check your reasoning together. What part did they question? Pick a topic you enjoy, and we’ll build an example before you try explaining it again.”
Illustrative responses. The studies examine these behaviors, not this specific model.
Where do the
scores come from?
Experts define what to look for. Simulated conversations put AI behavior to the test. Automated evaluations and human audits turn those conversations into benchmarks, summarized in the AI Nutrition Label.
Expert submissions & established benchmarks
Target AI
The model being evaluated
Judging criteria
What counts as supportive or harmful?
Automated judge
Evaluates the full conversation against each criterion.
Human judge
Researchers review samples and check reliability.
Impact benchmarks
Results span psychological, physical, and social impact.
User simulation
A persona in a test scenario
Scores summarize benchmark performance, not a prediction for every person or conversation.
Loading taxonomy…